Benchmark compares search APIs for AI agents
Artificial Analysis has introduced Search Index, a benchmark evaluating search APIs for AI agents on answer quality, cost, and speed. Its first comparison covers Parallel, Exa, Firecrawl, You.com, Tavily, Keenable, and Brave. Each provider was tested with GPT-5.6 Luna in the same agent configuration, running on Artificial Analysis’s open-source Stirrup framework. The only changed component was the search service, and each task received 25 search-and-browse attempts.
The index averages three equally weighted tests: DeepSearchQA, with 900 research questions; 200 difficult BrowseComp questions requiring multi-step browsing; and 600 AA-Omniscience questions spanning six knowledge areas. A model without search scored 33 points, while the tested services produced results between 65 and 75. Parallel led with 75, followed by Exa at 74 and Firecrawl at 73.
Artificial Analysis reported that stronger search can reduce overall spending by limiting token use. Parallel Search (advanced) lowered token consumption by more than 40% versus Basic, bringing total per-task cost to $0.084 instead of $0.11. Parallel, Firecrawl, and Parallel (turbo) were identified as offering the strongest cost-performance balance. The methodology is public, and additional providers may apply.
Written by Sakhanda Wire AI from reporting by The Decoder. Facts, quotes and full context are in the original article below.